MCP — the Model Context Protocol — is a standard way for an AI assistant like ChatGPT or Claude to use tools outside the chat. For notes, it means your assistant can, with your permission, search your Thoughts library, read a note in full, and save new notes into your projects. You control the connection with a sign-in and consent screen, and you can revoke it at any time. This guide explains the idea without jargon, then lists exactly what Thoughts does and does not allow through MCP.
The thirty-second version
Think of MCP as a well-labelled set of buttons an AI can press. A notes app exposes buttons like “search thoughts” and “save a thought”. When you ask a question, the assistant decides which button helps, presses it, and uses the result in its answer. The app — not the AI — decides which buttons exist and what each one may touch. No button, no capability: if there is no “delete” button, no amount of clever prompting can delete a note.
That last point is why MCP is reassuring for something as personal as notes. Access is not “the AI can browse my phone”. It is a short, auditable list of operations, each with a defined input and output, running under your account permissions.
How a connection is authorised
You add the app's MCP server address in your assistant's settings — for Thoughts that is https://thoughtsai.cc/mcp — and then sign in as yourself. A consent screen states, in plain language, what the assistant will be able to do. Approving it issues a revocable authorisation tied to your account, not a password handed to the AI. Every request the assistant makes is checked against the same project memberships you have in the app: if you cannot see a shared project, neither can it; if a collaborator removes you, access ends for both of you.
Because the authorisation is yours, revoking is simple and immediate. In Thoughts you do that from Connected apps; you can also remove the connector in ChatGPT or Claude. Step-by-step setup lives in the ChatGPT guide and the Claude guide.
The eleven things Thoughts allows
Thoughts exposes eleven tools: nine that read, two that create. In plain English, your assistant can:
- Confirm which account it is connected as.
- List your projects — yours, shared with you, or shared by you — and read a project's summary.
- List recent thoughts, filtered by project, author or last editor.
- Search thought titles, full text and tags for the words you give it.
- Read one thought completely, paging through long text rather than guessing from a preview.
- List collaborators you share projects with.
- Show recent activity — the latest state of things that changed in a window such as the last seven days.
- Show your invitation and shared-thought notification summaries.
- Create a new thought in a synced project, when you ask.
- Create a new cloud-synced project, when you ask.
That set is deliberately shaped around real sentences people say: “show my recent thoughts”, “find my pricing note in Launch”, “what changed this week?”, “save these next steps”. Asking AI about your own notes turns each into a worked example.
What stays off-limits — by design
The boundary matters as much as the features. Through MCP, an assistant cannot:
- Read local-only projects, locked thoughts, deleted items or anything not yet synced. Sync is your opt-in, per project.
- Access voice recordings, attachment files or drafts. It reads thought text only.
- Edit or delete existing thoughts or projects. Creation only — corrections are yours to make in the app.
- Invite people, accept invitations or change sharing. Reading a notification never marks it read or accepts anything.
- See full edit history. Activity shows the latest visible state in a window, not an audit log.
Search deserves one honest footnote: it matches the literal words you give it across titles, text and tags. It does not do semantic “you probably meant” matching. Predictable beats magical when the question is “did my AI really read my note?”
Why not just copy-paste or use an export?
You can, and for a one-off snippet that is fine. MCP earns its place when the interaction is repeated and two-way: the assistant finds the right note itself, reads all of it, reasons over several notes at once, and files the result back in the right project. Copy-paste makes you the courier in both directions and drops the filing step entirely — which is how useful answers end up lost in chat sidebars. Saving conversations as reusable notes compares the workflows honestly.
One server, every assistant
MCP is a protocol, not a ChatGPT feature or a Claude feature. The same Thoughts server works in both, which means your library does not fragment by assistant: save from Claude today, search from ChatGPT tomorrow. One second brain for ChatGPT and Claude explores why that matters more than any single model's memory feature.
When you are ready, the connection guide has the URL, the current steps for each assistant, and starter questions to verify everything works.
Frequently asked questions
What is MCP in simple terms?
MCP is a standard way for AI assistants to use tools outside the chat. An app exposes a fixed list of operations, like search or save, and the assistant can only do those things, under your account permissions, after you approve the connection.
Is MCP safe to use with my notes?
With Thoughts, access is limited to a defined tool list, authorised through sign-in and consent, checked against your own project memberships on every request, and revocable at any time. Local-only and locked content is never exposed.
Can an AI delete or edit my notes through MCP?
Not in Thoughts. The connector offers reading tools plus creating a new thought or project when you ask. There are no edit, delete or invitation-management tools.
Does MCP work with both ChatGPT and Claude?
Yes. MCP is a shared protocol. Thoughts uses the same server address for both assistants, so your synced library is the same whichever one you are talking to.